How Google's $1.50 Model Outscored Its Bigger One (Gemini 4 Hydrogen)
Google's new fast model just outscored the bigger model it sells beside it.
Gemini 4 Hydrogen launched November 12. On the Artificial Analysis Intelligence Index it scored 51. Gemini 4 Argon scored 50. Hydrogen is priced at $1.50 per million input tokens, and Artificial Analysis puts its cost per task at about 19% of Argon's.
The score matters less than how Google is using it:
- It built a working storefront in one shot (Google's claim).
- An agent called 3D Medic opens a broken 3D file, finds the errors and patches them.
- It sorts incoming bug reports, rates each one's complexity and routes the hard ones to Carbon, Google's top model.
- A new mode, Orchestrator, lets Carbon write the plan while Hydrogen writes and tests the code. Google says that used 24% less quota and finished 19% faster.
It does not win everywhere. On Terminal-Bench 4.0, Artificial Analysis has it at 54% and Claude Sonnet 5.5 at 64%.
The takeaway for anyone running agents: stop sending every step to your most expensive model. Let the big model plan. Let the fast one do the edits, the tests and the checks. Then compare the bill after a week.
Comment OPTIMIZE and I'll send you our tips for routing your tokens to the best model for each type of task.
Clips and charts: Google (blog.google, @GeminiApp, @GoogleAI, @googledevs, @Google and @GoogleDeepMind on X), Google Antigravity (antigravity.google) and Artificial Analysis (@ArtificialAnlys on X).






Comments1
Facts. For a few months, I've been using orchestration through Kun Chen's Firstmate. With the recent boom of new Gemini models, I've been loving how Google is handling orchestration so far on Antigravity. Overall, it has been quite resourceful.
1 comment
Facts. For a few months, I've been using orchestration through Kun Chen's Firstmate. With the recent boom of new Gemini models, I've been loving how Google is handling orchestration so far on Antigravity. Overall, it has been quite resourceful.